Load balancing and rate limiting based algorithms for improving cloud computing performance

Abstract

The cloud computing paradigm has recently become an increasing popular method for the delivery of internet services. While the increased utilisation of the cloud has made it easier for cloud operators to recover their capital investment further work can be done to maximise profits. The cost of constructing a cloud is considerable (hundreds of millions of Euro) and the operating cost is also significant (tens of millions of Euro annually). Cloud operators can attempt to ensure that their capital investments are recovered by lowering operating costs and using mechanisms which endeavour to maintain high utilisation. In this dissertation we begin by examining methods which cloud operators can use to control bandwidth utilisation to ensure high utilisation. Fixed cost pricing is more desirable to enterprises and we examine the use of distributed rate limiting to achieve this in the cloud. In addition, some clouds are public environments and as such mechanisms that ensure that each user receives a fair share of the bandwidth available are useful in maintaining quality of service levels. We propose the use of a dropping mechanism to achieve this and compare it with the established token bucket mechanism. We then shift our attention to investigate methods which cloud operators can use to lower the operational costs. Firstly we propose algorithms to lower cooling costs while ensuring all demand is serviced. Data centres generate a lot of heat and this must be removed to prevent damage to equipment. Some data centres are excessively cooled to cater for worst-case airflows. By equalising the inlet temperature of server racks within a data centre we attempt to lower the cooling cost. Secondly we examine algorithms to lower carbon emissions while maintaining a reasonable Quality of Service (QoS) level. The carbon intensity of electricity suppliers, which is the carbon emitted to produce a given amount of electricity, varies in different geographical regions. Our work formulates the operation of the cloud as an optimisation problem and applies the subgradient method to minimise the combination of average service request time and carbon emissions. Finally we propose an algorithm which uses Voronoi partitions to minimise a function which encompasses the electricity cost, carbon emissions and average service request time. Both the electricity cost and carbon emissions are affected by the cooling design used at the data centre and we also incorporate this in our simulations in this section. We gather carbon intensity, electricity price and latency data for a cloud which has data centres in USA and Europe (similar to a popular commercial cloud). Our work then utilises this data and examines how this algorithm can be used to achieve a variety of goals such as minimising carbon emissions, electricity cost and average service request time.